Video Objects Extraction And Trajectory Tracking

Since the advent of video survillence, it developed in three stages：traditionalsimulation video monitor system, digital video monitor system, network videomonitor system. As the rapid development of computer technology, people eager torealize intelligence in the field of video survillence, to relieve the pressure of artificialmonitoring, it propel the development of the video survillence to its next stage:intelligent video survillence.This paper mainly studies objects extraction and trajectory tracking in the videosurvillence environment. Through analysis and comparing several algorithms, wechoose a reliable algorithm. At the stage of extraction, we use an algorithm based onaccumulated-frame-subtraction and twice-frame-subtraction. First,to avoid incompleteobject extraction caused by moving slowly, we accumulate the forward and backwardframe difference images. At the same time, to avoid redundant, we control theaccumulated number through comparing the variance of the accumulated framedifference image and the given threshold. Then remove the false target area caused byobject movement through taking the intersection of the forward and backward framedifference images. At the stage of tracking, we rely on the objects’ position, shape,and brightness statistics feature to avoid single feature easily affected. It increases therobustness of the algorithm.Occlusion is a big problem that affected the accuracy and stability in tracking. Tohandling occlusion, an algorithm based on weighted luminance histogram is proposed.First, using the prediction of Kalman filter to deal with the objects that failed inmatching, we can get the centroid position in current frame, and judging if they are inocclusion or not by the distance between the centroid. If there are two objects inocclusion, distinguish the active object and the passive object through the position inspace. Then calculate the two objects’ weighted luminance histograms. Set smallervalue to the part close to the occlusion edge that easily affected, and set larger value tothe opposite part. Find an fail-matched object in current frame whose width is largerthan the active object or the passive object, divide it into two areas, calculate theweighted luminance histograms. We can successfully track the objects underocclusion through calculating the similarity of the weighted luminance histogramsbetween the objects and areas.The fail-matched objects which are not be covered by each other would bejudged to be covered by the obstacles in the background. Calculate the weightedluminance histograms from both directions and give the similarity between thefail-matched objects in prior frame and current frame. If the handling fails, the fail-matched objects in prior frame are missing, the fail-matched objects in currentframe are newly appeared.At the beginning of tracking, establish a set for the appearing objects, update theinformation of the objects when the tracking was completed in the current frame tokeep coherence. And, to the missing targets, they may appear in the later frames onceagain, keep the information temporarily. If they were missing in the continuouspassed10frames, they will be judged to the forever-missing objects, and clear theinformation about the objects from the set.The experiment results indicate that the algorithm in this paper can do accuratelyin extraction and tracking. When occlusion happens, the handling occlusion algorithmproposed in this paper can make most of the useful information, keep the continuityand the stability of the tracking system.